How can you say that when we have
absolutely no information whatsoever on what filesystem he is using?
what does filesystem have to do with ext-todo? ext-todo is for high
injection rates, where qmail-send will spend more time pre-processing
messages than delivering them.
ext-todo needs to be able to read in the stuff to do its job. If the
filesystem cannot handle large numbers of files in a directory, it will
bog down trying to process todo.
Now, I admit that there is some missing information before we can
properly diagnose silly qmail syndrome... we haven't seen his qmail-send
logs, so we aren't sure if he's maxing out his remote concurrency (if he
is, he's not suffering, if he isn't, he is suffering), but ext-todo
still doesn't have anything to do with filesystem.
He is not maxing his concurrency all the time but the symptoms already
indicated silly qmail syndrome and his reply shows that i hit the
jackpot. The graph from qmail-mrtg did not quite demonstrate it but it
was enough that it was plausible although not a constant problem. That
is why I also said his problem seems to be seasonal and if he does not
mind a large queue now and then, he really does not have to do anything.
What do you mean ext-todo has nothing to with the filesystem? Everything
has to do with the filesystem. Why do you think people run tests on what
filesystem is best for a certain application? Most especially for an i/o
intensive applications such as mtas.
Did you miss those emails asking what is the best split number? Why do
they ask that? Due to filesystem behaviour. Filesystems with directory
index support could use a split of 1 instead of 23 while others on
filesystems without directory support might need 59 or higher due to the
volume they deal with. And you want to say ext-todo has nothing to do
with the filesystem when it has to access it to get its job done and the
optimal configuration of qmail is influenced by the filesystem it uses?
What do you know of optimizing a qmail queue to deliver? Have you ever
tried to split the queue across two or more filesystems sitting on their
own disks? Have you ever done a setup where concurrencyremote is over a
thousand just to make sure dead mail does not eat up all available
remote slots on a single box that handles over a million remote
deliveries daily?
|